A Hybrid Approach for Detecting Bugs in HPC Workloads.

MPI programs represent a major type of workloads running on parallel and distributed systems: tightly coupled high performance computing (HPC) workloads which use MPI to communicate between processes and instances. One of the major challenges for MPI programs is bug detection. Traditional approaches for diagnosing MPI bugs attempt to reproduce the exact execution schedule which reveals the bug, resulting in high run-time overhead. In this paper, we present our work in identifying bugs in MPI programs using a hybrid approach that leverages both static and dynamic analysis, and detects bugs at both compile time and run time. The static approach detects buggy patterns by analyzing the Intermediate Representation (IR) file generated using the LLVM compiler. The dynamic approach takes user defined rules, verifies them at runtime, and detects any violations which could be caused by a bug. To evaluate our approach, we have carried out experiments to detect various bugs in different benchmarks such as CombBLAS, OpenFFT, and NAS Parallel benchmarks. Our results show that the hybrid approach is effective in detecting bugs at both compile time and runtime.